Abstract
The purpose of developing a cost-effective smart walking stick for the visually impaired and the need increased with a steady increase in their population worldwide. The smart stick presented in this paper uses artificial intelligence along with ultrasonic sensors to help the visually disabled people to navigate through the environment independently. Image recognition, Collision detection and obstacle detection are the tasks performed on a raspberry pi having high processing power. The task of obstacle detection and obstacle avoidance makes use of ultrasonic sensors to alert the user of the obstacles appearing in his path. The smart walking stick also demonstrated the important characteristics of affordability, mobility and high performance.
Introduction
Visually impaired individuals face significant challenges in navigating unfamiliar environments safely and independently. Traditional aids, such as the white cane, provide limited range and no contextual information about the surroundings. This paper introduces an AI-enhanced Electronic Travel Aid (ETA) that bridges this gap by combining local sensor-based collision avoidance with cloud-based computer vision for scene understanding.
Problem Statement
With an estimated 45 to 50 million blind people globally, many of whom live in low-income regions, there is a dire need for assistive technology that is both high-performing and affordable. Existing high-tech solutions often suffer from high latency, bulky form factors, or excessive costs. Furthermore, many systems rely on audio feedback that can mask environmental sounds crucial for safety. This research seeks to create a mobile, affordable, and high-performance device that balances local "reflexive" sensing with sophisticated "cognitive" scene understanding.
System Architecture and Methods
The proposed system is centered on a Raspberry Pi 3 Model B+ integrated into a standard walking stick. The architecture follows a hybrid processing model to balance speed and intelligence.
Local Obstacle Detection
For immediate safety, the system utilizes an HC-SR04 ultrasonic sensor. This sensor continuously monitors the path ahead. When an object is detected within a predefined "safe zone" (typically 4 feet), the Raspberry Pi triggers haptic feedback via a vibration motor and audible beeps. This local processing ensures near-zero latency for collision warnings, which is critical for real-time navigation.
Cloud-Based Scene Understanding
To provide high-level context, a Pi Camera captures images at 15-second intervals. These images are transmitted via a smartphone hotspot to the Microsoft Azure Computer Vision API. The API analyzes the scene, generating descriptive tags and captions (e.g., "a dog looking at the camera"). These descriptions are then converted to speech using Google TalkBack and delivered to the user through earphones.
Key Insights
- Hybrid Processing: By separating "reflexive" tasks (collision detection) from "cognitive" tasks (image recognition), the system maintains safety while providing rich environmental data.
- Leveraging AI-as-a-Service: Utilizing pre-trained models from Microsoft Cognitive Services allows the device to remain lightweight and cost-effective without sacrificing recognition accuracy.
- Multimodal Feedback: The use of both vibration and audio alerts helps prevent "auditory masking," ensuring the user remains aware of their physical surroundings while receiving digital assistance.
- Affordability: By using off-the-shelf components like the Raspberry Pi and a standard 10,000 mAh power bank, the authors demonstrate that sophisticated assistive tech can be built for a fraction of the cost of specialized medical devices.
Performance Results
The system was evaluated based on its response time across various network conditions. The results demonstrate that the hybrid cloud-edge approach significantly reduces latency compared to traditional mobile-app-based architectures.
As shown in the performance data, the system achieves a processing time of approximately 4-5 seconds on a 4G network, a substantial improvement over the 18-20 seconds seen in older systems. This reduction in latency is critical for providing timely information to a moving user. Additionally, the ultrasonic sensors proved reliable in detecting common hazards such as potholes and bumps, providing a buffer that allowed users to alter their route in real-time.
Conclusion and Future Work
The AI-powered smart walking stick offers a promising, affordable solution for independent navigation. While currently dependent on internet connectivity for its vision features, the system demonstrates high performance in obstacle avoidance and scene description. Future iterations will focus on improving ergonomics, reducing power consumption, and exploring offline edge-AI capabilities to ensure functionality in areas with poor network coverage.
Summary updated